Support for recurrent_v2 layers for TF 2.0

[cjermain]

In #607, it was observed that TensorFlow versions before 2.0 can have problems converting TF 2.0 compatible models because the classes are expected to be the previous API. For example in TF 1.14 the v2 compatible classes are present in tensorflow.keras.layers.recurrent_v2, but do not match tensorflow.keras.layers. This PR resolves #607 by adding support for these layers. For versions of TensorFlow that are above 2.0, this is not a problem because tensorflow.keras.layers uses the v2 layers.

This strategy is tested by expanding the RNN_CLASSES list in the tests, so that recurrent_v2 classes are tested if present.

[cjermain]

@wenbingl, it looks like the linux-conda-ci build is still waiting to start. Any ideas on how to resolve it?

[wenbingl]

@wenbingl, it looks like the linux-conda-ci build is still waiting to start. Any ideas on how to resolve it?

this is the recent random failure of azp, we don't have a good way to fix it.

And recently, ONNX sig added a DCO requirement which you needs sign your commit. After you sign the commit, I can try to manually run the checker and forcedly merge it if the azp issue still exists.

[cjermain]

Ok, got it. I re-committed and force-pushed the change since it was buried under a few merges.

[jiafatom]

This PR fails on linux-tf-keras CI build for tensorflow==1.15.0. The fix is here.